Should I vibe code
Mac app for local Whisper transcription, subtitles, and batch processing
whisper.cpp plus a drop zone is the whole app. The only real question is who else was on the recording.
?
Their verdict, the MacWhisper Pro price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-05.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
This is the shape of project the site exists to say yes to. whisper.cpp ships Metal and Core ML builds, the models are a download, and a Mac app that takes a dropped file and writes an SRT beside it is genuinely an evening — no account, no per-minute billing, no vendor, and an exit that consists of dragging an app to the trash. Two honest caveats, neither of which moves the verdict. The first is a boundary that is easy to cross by accident: wire in a hosted API because the large model is slow and you have quietly built a cloud transcription product, with every question this site asks of one. The second is not about your code at all — the system-audio capture that makes meeting transcription work is a recorder, and whether you may point it at a call with other people on it is a question about consent, not about software.
What actually breaks
not "if". the specific failures.
- Whisper on long files, which loops and hallucinates over silence and music — and the failure arrives as confident, fluent English that nobody said, in the middle of an otherwise accurate transcript
- Subtitle timing, which is the whole difference between a transcript and a usable SRT, and is not what the model optimises for
- Speaker diarization, a separate model and a separate pipeline, and the place where "beta" is doing genuine work in every product that ships it
- Long batch runs: thermal throttling, the machine sleeping, the app being quit at hour three, and no resume because nothing checkpointed
- Model files, which are multi-gigabyte downloads that nobody checksums, with quantisation levels that trade accuracy in ways you will not notice until a name comes out wrong
- The output folder, which quietly accumulates transcripts of everything you have ever recorded, in plain text, in a directory your backup tool uploads without asking
- Microphone and system-audio permissions, which macOS resets on major updates, turning a working recorder into a silent one
- Proper nouns and jargon — wrong in precisely the sentences you were transcribing for
Is that you?
the verdict is a default, not a law
- Models run locally and nothing is uploaded, which is the entire point of choosing this over a hosted transcriber
- The audio is yours: your voice memos, your recordings, files you already had the right to hold
- You read the transcript before quoting it, because Whisper is fluent about things that were never said
- The transcripts live somewhere you chose deliberately, with an eye on what is in them
- It records calls other people are on without telling them — that is a consent question in most of Europe and in every all-party-consent US state, and it does not care that your tool is local
- You added a hosted API for the big model and still describe the app as private
- Transcripts of client, legal, medical or HR conversations land in a folder that syncs to a cloud drive by default
- You are treating the output as a record of what was said rather than a draft of it
If you build it anyway
the checklist, then the prompt that enforces it
- Keep everything local and make that verifiable: no network calls at all in the transcription path, and say so in the README. If a hosted model is ever added, it should be a separate, clearly labelled mode.
- Verify model checksums on download and record which model and quantisation produced each transcript, because that is the only way to explain a bad result later.
- Checkpoint batch jobs per file and support resume. A queue that loses four hours of work because the lid closed is the most common self-inflicted wound here.
- Write transcripts to a directory you chose deliberately, outside iCloud Drive and Dropbox, and offer a one-command purge.
- If you build recording, announce it: a persistent, visible indicator while capture is running, and no automatic start on meeting detection.
- Handle silence explicitly — voice-activity detection before the model — because that is where hallucinated text comes from.
- Export SRT and VTT with timings you can edit outside the app, and keep the plain text separate from the timed version.
- Check the microphone and system-audio permissions on every launch and fail loudly, since the failure mode is a recording of nothing.
I am building a Mac transcription app around local Whisper models: drop a file,
get text and subtitles, with batch processing and optional recording. Keep it
genuinely local, and treat recording as the sensitive part.
1. No network in the transcription path. Bind whisper.cpp or a local Core ML
model and make an offline run the default and the only default. If a feature
would need a hosted API, stop and tell me before writing it.
2. Verify the checksum of every downloaded model and store which model and
quantisation produced each transcript alongside the output.
3. Run voice-activity detection before the model and skip silent regions.
Whisper's hallucinations come from silence and music, and this is the cheapest
mitigation there is.
4. Build batch as a resumable queue: checkpoint per file, survive a quit, never
restart four hours of work because the lid closed.
5. Write outputs to a directory I choose, defaulting outside iCloud Drive and
Dropbox, and print the path on first run. Give me one command that deletes
every transcript.
6. Ship SRT and VTT export with editable timings before any styling or burn-in,
and keep the plain-text transcript as a separate artefact.
7. If you add audio capture: a persistent visible indicator while recording, no
automatic start on meeting detection, and a first-run explanation that
recording other people may require their consent where I live.
8. Check microphone and system-audio permissions on every launch and show a
persistent warning when they are missing. A silent recording is the worst
outcome, because it is discovered later.
9. Surface low-confidence segments in the UI rather than hiding them. I need to
know which sentences to re-listen to.
10. Out of scope until I ask: cloud sync, summarisation by a hosted model,
uploading audio anywhere, and sharing transcripts with anyone else.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
Buy for the polish, not the transcription. A one-time €65 is a batch queue that resumes, diarization that has been tuned, subtitle timing that survives a real editor, a corrections dictionary for the proper nouns you keep getting wrong, and someone else keeping up with macOS permissions and new model releases. The engine underneath is the same open-source thing you would have used. If your need is "turn this folder of voice memos into text", build it — that version really is an evening and it really is enough.
There is barely an exit to plan. The models are files, the outputs are text, and the app is a wrapper — delete it and you keep every transcript you ever made. The only thing worth doing deliberately is keeping SRT and plain text side by side, so nothing you produced depends on your own project file format. If you later move to MacWhisper itself, or to Buzz, they read the same models and produce the same formats.
The C/C++ Whisper implementation with Metal and Core ML support that almost every local Mac transcriber is built on.
Cross-platform open-source GUI for offline Whisper transcription and subtitle export.
Questions
Is it really all local?
It can be, and that is the whole reason this entry is green — whisper.cpp runs the model on your machine with Metal, and the only thing that ever touches the network is downloading the weights once. The boundary is easy to cross by accident, though: reaching for a hosted API because large-v3 is slow on your laptop turns a private tool into a service that receives your audio. If you do that, at least be honest in your own README about which mode is which.
Does recording a meeting change the verdict?
It changes what you have to do, not what you have to build. Capturing system audio from a call is legally a recording, and consent rules follow the participants rather than the software — several US states require every party to agree, and in the EU you need a lawful basis under GDPR Art. 6. That is why regulatoryExposure is 3 rather than 0. If your product becomes "a bot that joins meetings and distributes summaries", you are in the Granola and Fireflies territory instead, and those entries are considerably less relaxed.
Why does Whisper invent text, and how bad is it?
The model was trained to produce fluent transcription and has no way to say "there was nothing here", so silence, background music and long pauses can produce plausible sentences that were never spoken — sometimes recycled from the training data. It is not frequent and it is not random-looking, which is what makes it dangerous: the hallucinated line reads exactly like the rest of the transcript. Voice-activity detection before the model removes most of it, and reading before quoting removes the rest.
- RCFP reporters' recording guide (US, per-state consent law)
- GDPR Art. 6 — lawfulness of processing
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice